PREDICTION FOR HEAVY RAINFALL USING NEURAL NETWORKS

本文的目的是建立一种对一定降雨的风险度进行预测的方法，以此作为应对降雨引起的泥沙灾害的对策。在此基础上，定义了集中时间和临界降雨量，构建了基于神经网络的预报系统。该方法的主要步骤如下：根据一个降雨时间序列计算各时刻的累计降雨量。给系统提供归一化后的雨型和发生0.99或不发生灾害的示教信号。通过反向传播学习对神经网络进行优化后，系统自动计算出新雨型的风险度从0到1，并将该预测系统应用于三种类型的实际案例，结果表明，该系统具有较高的判断精度，适用于云顶火山泥石流发生的预测。对9119号台风后奇库果河流域的次生灾害，发现台风后临界降雨量的变化。此外，还得到了集中时间和临界降雨量。阐明了前期降雨对鹿儿岛市泥沙灾害发生的影响，该系统是泥沙灾害预报和泥沙灾害发生判据的良好工具。因此，利用神经网络进行降雨率预报的系统在未来似乎会有更大的发展。

The goal herein is to establish a prediction method for a degree of risk by certain rainfalls , as a countermeasure against sediment disaster which caused by rain. Thus, the time of concentration and critical rainfall are defined, A prediction system using neural network is constructed. It is expected that the neural network can learn a general rule from examples.The main procedure of the method are as follows ; cumulative rainfalls on various times are calculated from a rainfall time series. The rainfalls pattern which is normalized and teach signal of occurrence 0.99 or non-occurrence 0.01 of disaster are given to the system. After the neural network is optimized by back-propagation learning, the system automatically calculates the degree of risk from 0 to 1 for new rainfalls pattern.The prediction system is applied to three types of real cases and the results are as follows ; It is applicable for predicting the occurrence of debris flows at Unzen volcano, because the system has high accuracy in judgment. On secondary disaster after typhoon No.9119 at the Chikugo river basin, change in critical rainfall after the typhoon is recognized. Moreover the time of concentration and critical rainfall are found. It is clarified that the antecedent rainfall affects the occurrence of sediment disasters in Kagoshima city.This system is good tool for prediction of sediment disaster and to estimate the occurrence criteria. Therefore, the prediction system using the neural network for the prediction of the rainfall rate seems to be developing more in the future.